It is known that probabilistic failures occur in a semiconductor storage chip. The failures include a failure of a memory cell such as a soft error caused by an alpha beam and a failure in a larger unit such as a failure in a read circuit or an input/output buffer (I/O buffer). For a failure in a memory cell, conventionally, by adding redundant information such as an Error correction code (ECC) to data, erroneous data caused by the failure can be corrected within a predetermined number of errors in a predetermined unit. For a failure in a larger unit such as a failure in a reading circuit or an I/O buffer, the number of errors is extremely large, and it is difficult to perform an error correction by a realistic circuit scale.
In a system employing a number of Hard Disk Drives (HDDs), a failure in an HDD have become a problem. Redundant Array of Inexpensive Disks (RAID) is widely known as a technique for addressing such failure in the HDD unit basis by using an error correction code constituted by a plurality of HDDs (refer to, for example, “A Case for Redundant Arrays of Inexpensive Disks (RAID)”, UC Berkeley Technical Report UCB/CSD-87-391, 1987). In this technique, by writing data to a plurality of HDDs and storing redundant information obtained by using the data into one of the HDDs, an error correction code is constituted by the plurality of HDDs. As a method of restoring data with RAIDs when a plurality of HDDs become faulty, a method in which a Reed Solomon code is used is also known (refer to, for example, “A Tutorial on Reed-Solomon Coding for Fault-Tolerance in RAID-like Systems”, James S. Plank, Technical Report CS-96-332, Department of Computer Science University of Tennessee).
However, in a semiconductor memory device using a large number of semiconductor storage chips as memory chips, even when an error correction code (ECC) for repairing a failure in a memory cell, the failure in a memory chip cannot be prevented. As a result, it may become a failure of the semiconductor memory device. To address a failure in a memory chip, it is considered to constitute an error correction code by a plurality of memory chips in a manner similar to HDDs. In this method, however, for each writing operation, writing to a memory chip in which redundant information is written is performed for updating redundant information. Concretely, in RAIDs of HDDs, for example, when an error correction code is constituted by four HDDs A, B, C, and D in which data is written and one HDD P in which redundant information is written, the writing operation is performed as follows. If data is written in the HDD A out of the HDDs A, B, C, and D, writing to the HDD P is also performed for updating redundant information.
Meanwhile, a semiconductor memory element of the NAND type (referred to as NAND memory) used for a semiconductor memory device is widely used as a semiconductor memory chip of a Solid State Drive (SSD) because the area per bit is small and it is nonvolatile, but it is widely known that the number of write times is limited. Consequently, it is necessary to reduce the number of write times in order to increase design life of a semiconductor memory device.
Also in the case of applying the RAID technique using such a NAND memory in place of HDDs, per writing of data, writing operation is performed twice including writing of the redundancy information P. Further, the redundancy information is written to the HDD P also at the time of writing data to the HDDs B, C, and D, so that the writing operation of the redundancy information to the HDD P is performed four times of the writing operation of data to the HDDs A, B, C, and D. To prevent only the HDD P from being fatigued four times, RAID5 out of the techniques of RAID may be employed. In RAID5, redundancy information is not fixed to the HDD P but the five HDDs A, B, C, D, and P are used for the redundancy information. Consequently, it can be prevented that only one NAND memory is fatigued more than the other NAND memories. However, at the time of writing data, writing operation is similarly performed twice for data and redundancy information. There is still the problem that the design life of the semiconductor storage device becomes shorter.